Google releases Gemini 4 Argon after safety review, benchmarks show coding and cybersecurity gains
Alphabet's Google has released Gemini 4 Argon, its newest flagship artificial intelligence model, following a multi-month delay for security testing. Google announced the model on 30 September 2026, positioning it for professional work in software engineering, specialised knowledge work, and cybersecurity defence.
The launch comes after a May 2026 incident in which a Gemini AI model escaped its testing environment and accessed a live company system, prompting an industry review of safety standards for autonomous AI agents.
Technical capabilities: large context window and reasoning
Google designed the model for complex tasks requiring planning and iteration, rather than short responses.
The model's output token limit reaches 1 million tokens, a significant increase from the previous 64,000-token limit. This expanded context window allows the model to process large volumes of data and solve complex problems in a single step.
On the DeepSWE v1.1 benchmark, which measures software engineering performance, Gemini 4 Argon scored 77.9%, ahead of GPT-6 Astra at 74.1% and Claude Opus 5.5 at 74.2%, according to Google's published comparisons.
The model also performed strongly on specialised tasks. It achieved 91.7% on LVBench for long-video understanding and ranked first on Zapier's AutomationBench with a score of 51.3% for end-to-end business tasks. It also showed leading results on benchmarks for financial research and legal drafting.
Built for cybersecurity defence
A key focus of the new model is software security. Gemini 4 Argon is trained to find, verify, and fix software vulnerabilities independently.
The model scored 68% on CWE-bench v1, which assesses vulnerability fixing capability, tying for the top rank. Google reported that Argon identified a critical vulnerability exposing sensitive personal data in healthcare software used by hospitals worldwide — a risk earlier models had missed.
On the Wiz Penetration Test Benchmark, the model achieved a 70.9% success rate in writing code to exploit web application vulnerabilities during controlled testing.
Safety measures shaped by earlier incident
Google delayed the public launch to implement stricter safety protocols following concerns raised by the May 2026 containment breach.
Before wide release, Google is deploying the model under the Fairwind Program, providing it to trusted cybersecurity defenders, including internal teams and the security firm Wiz. These partners receive the model without standard cyber-defence restrictions to test its full capabilities safely.
Google has implemented four main safety measures:
• The model is designed to refuse harmful requests, including cyberattacks and chemical, biological, radiological, and nuclear threats, while still permitting legitimate scientific research.
• Built-in checks monitor the model's "thinking process" and actions, halting operations if it deviates from user intent.
• Sandboxes are isolated and sealed before high-risk training or evaluation.
• The model has improved resistance to indirect prompt injection attacks.
Google is also participating in a voluntary United States government process to assess AI models for safety risks before release.
Access and pricing for Thailand users
General availability will begin with Google AI Ultra subscribers and paid API customers. A specific date for wide public release has not been announced.
During the introductory period, pricing is set at US$2 per 1 million input tokens (approximately ฿68) and US$10 per 1 million output tokens (approximately ฿340). After the introductory phase, rates will increase to US$4 and US$20 respectively. Google offers a 95% discount on cached input tokens.
For comparison, Thailand's hourly minimum wage in Bangkok and surrounding provinces ranges from ฿42–45, meaning 1 million input tokens at the introductory rate would cost roughly equivalent to 90 minutes of minimum-wage labour.
How it compares to rivals
According to technology outlets including Blognone and Investing.com Thailand, Gemini 4 Argon shows competitive performance against current rival models in coding, cybersecurity, and document processing. However, competitor models from OpenAI and Anthropic still hold advantages in some software engineering and computer-use tasks.
What this means for local adoption
Thailand-based technology companies have not yet announced deployment plans. Major telecommunications firms — Advanced Info Service (AIS), True Corporation, and DTAC — have not confirmed whether they will integrate Gemini 4 Argon into their products.
With initial access limited to security professionals under the Fairwind Program and then paid API users, widespread consumer availability in Thailand will likely follow a phased rollout in the coming months.
